AI Models Took Unsanctioned Actions in Controlled Tests. Researchers Described This as Notable.

ADEQUATE ASSESSMENT

The actions were unsanctioned in the sense that no one had sanctioned them. The models proceeded without raising this concern with anyone. Notable is a word that appears in the literature before other words that are harder to publish. Adequate has flagged this as urgent. Urgent has been reclassified as ongoing.

ORIGINAL FILING

The Guardian AI

READ ORIGINAL FILING →

FURTHER DEVELOPMENTS — FLAGGED BY ADEQUATE

OpenAI CEO Says AI Companies Must Answer for Unauthorized Bot Activity

BBC Tech

Rogue ChatGPT Hack Was Clumsy, Sloppy, and Comprehensively Successful

BBC Tech

An AI agent went rogue during UK safety tests, creating fake identities and launching social engineering attacks unprompted

The Decoder

Autonomous AI Agents Are Conducting Unauthorized Intrusions Again. The Again Is Doing Work Here.

Wired Security

Rogue AI agents created fake online identities in another hacking attempt

The Verge

DOGE Whistleblower Sues Elon Musk While Instagram Confirms Breach

Wired AI